Cost tracking for AI agents
Know what your
agent spends.
CostMeter sits between your AI agent and API providers. Every token counted, every dollar tracked, every budget enforced. One command to install, zero config to start.
# Install + auto-configure (detects OpenClaw, Claude Code)
curl -fsSL https://costmeter.io/install.sh | shFree to use. No account required. Runs on your machine.
The problem
AI agents make API calls you don't see. A coding session burns through tokens. A research task hits three providers. A scheduled job runs while you sleep.
By the time you check the dashboard, the bill is already there. CostMeter gives you visibility and control before things get expensive.
Works with your stack
CostMeter auto-detects your tools and configures itself. No YAML editing. No env var juggling.
OpenClaw
Auto-imports all provider keys and patches config
Claude Code
Detects credentials and patches settings.json
Any Agent
Works with any tool that supports ANTHROPIC_BASE_URL or OPENAI_BASE_URL
Key features
Real-time cost tracking
Every API call logged with provider, model, tokens in/out, latency, and dollar cost. Query by day, week, or month. Group by provider, service, model, tool, or project.
Budget enforcement
Set hard limits per provider, service, or project. Soft warnings at 80%, hard blocks at 100%. Atomic check-and-reserve prevents race conditions in concurrent agents.
Rate limit visibility
Captures rate limit headers from Anthropic and OpenAI responses. See remaining requests and tokens in real time. Know when you're about to hit the wall.
OAuth passthrough
Works with subscription plans (Claude Pro/Max) without touching auth. Your OAuth tokens pass through untouched. CostMeter tracks tokens and rate limits, not dollars.
Alerts where you are
Budget warnings via webhook, Slack, Discord, or piped through OpenClaw to WhatsApp/Telegram. With cooldown dedup so you don't get spammed.
Self-updating
Runs as a system service (systemd/launchd). Survives reboots, auto-restarts on crash. Weekly auto-update with SHA-256 checksum verification.
Built for agents
Not a dashboard you check after the fact. CostMeter is a proxy that sits in the hot path. SSE streaming passes through chunk-by-chunk with zero buffering. Cost is injected as an SSE comment so your agent can read it in real time.
# Agent → CostMeter → Anthropic (streaming)
data: {"type":"content_block_delta","delta":{"text":"Hello"}}
: costmeter_cost=0.000034
data: {"type":"content_block_delta","delta":{"text":" world"}}
: costmeter_cost=0.000067
data: {"type":"message_stop"}
Supported providers
+ any OpenAI-compatible endpoint via format: openai — Together, Fireworks, Ollama, etc.
Everything from the CLI
# What did I spend today?
$ costmeter spend --by provider
PROVIDER COST REQUESTS
anthropic (sub) 47
openai $2.34 142
google $0.85 170
TOTAL $3.19 + subscription
# How much headroom do I have?
$ costmeter limits
anthropic (claude-sonnet-4)
Requests: 950 / 1000 remaining (95%)
Tokens: 480K / 500K remaining (96%)
# Watch live
$ costmeter tail
# Full diagnostics
$ costmeter doctorYour data stays yours
CostMeter runs on your machine. API keys never leave your system. Prompts and responses are not stored — only metadata: provider, model, token counts, cost, latency. The SQLite database lives at ~/.costmeter/costmeter.db.
Stop guessing.
Start tracking.
One command. Auto-detects your setup. Running in 30 seconds.
curl -fsSL https://costmeter.io/install.sh | sh